A Deployment-Oriented Benchmark of Machine Learning Models for Agroclimatic Forecasting Under Degraded Sensor Data

This study evaluates the performance of classical machine learning models for one-step-ahead agroclimatic time-series forecasting under degraded sensor-data conditions. The motivation is the operation of IoT-based field monitoring systems, where measurements may be noisy, incomplete, temporally irregular, and constrained by limited local storage and computational resources. Using a real meteorological dataset collected by a field weather station in the Dnipro region of Ukraine, we compared twelve regression models: Ridge Regression, Random Forest, Extra Trees, Gradient Boosting, HistGradientBoosting, Support Vector Regression, Linear SVR, KNN, PLSRegression, ElasticNet, Lasso, and MultiTaskElasticNet. The models were evaluated under five controlled experimental scenarios: baseline data, missing values, additive noise, reduced training history, and combined noise–missingness degradation. The results indicate that Ridge Regression provides the highest accuracy under clean and mildly degraded conditions, whereas HistGradientBoosting is more stable under severe combined degradation. These findings support the use of deployment-oriented model selection for agroclimatic forecasting; however, the proposed edge/fog workflow should be interpreted as a conceptual deployment direction unless validated by direct measurements of latency, memory footprint, and energy consumption on representative hardware. The study therefore provides a benchmark for robustness-oriented model selection rather than a fully validated embedded deployment framework.